GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting

Yan et al. present GS-SLAM, one of the first systems to adopt 3D GaussianSplatting as the underlying scene representation for dense RGB-D SLAM. Earlier neural SLAM methods, such as Point-SLAM and Co-SLAM, rely on implicit features anchored in points or hash grids and reconstruct geometry via NeuralImplicitSurface rendering. Those representations couple ray sampling to MLP queries and impose a substantial cost per pixel, which limits incremental update rates. GS-SLAM instead uses Gaussians and a differentiable splatting rasteriser, exploiting the same speed advantage that 3DGS brings to offline reconstruction.

The system runs a two-track loop. Tracking estimates the new camera pose by minimising a photometric-and-depth re-rendering loss against the current Gaussian map. A coarse-to-fine selection of reliable Gaussians focuses pose optimisation on confidently observed regions, suppressing the influence of recently spawned or noisy primitives. Mapping, in turn, runs an adaptive expansion strategy that adds new Gaussians where the current map fails to explain incoming pixels and prunes Gaussians that drift away from the surface. This combination is essential for extending 3DGS — originally a batch reconstruction method — to incremental, scene-scale capture.

A practical advantage of the explicit primitive is that the system can visualise the current map at any time without an extra rendering pass through a separate decoder, which simplifies debugging and live monitoring of large captures. On Replica and TUM-RGBD, GS-SLAM achieves competitive tracking accuracy and state-of-the-art rendering quality among RealTimeRendering SLAM methods. It is contemporary with Gaussian Splatting SLAM by Matsuki et al., which tackles the harder monocular setting; together the two papers established explicit Gaussian primitives as a serious option for live dense SLAM and seeded a follow-on literature on Gaussian-based incremental reconstruction.